CN107437245B - High-speed railway contact net fault diagnosis method based on deep convolutional neural network - Google Patents

High-speed railway contact net fault diagnosis method based on deep convolutional neural network Download PDF

Info

Publication number
CN107437245B
CN107437245B CN201710493102.5A CN201710493102A CN107437245B CN 107437245 B CN107437245 B CN 107437245B CN 201710493102 A CN201710493102 A CN 201710493102A CN 107437245 B CN107437245 B CN 107437245B
Authority
CN
China
Prior art keywords
equipotential line
equipotential
picture
neural network
convolutional neural
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Expired - Fee Related
Application number
CN201710493102.5A
Other languages
Chinese (zh)
Other versions
CN107437245A (en
Inventor
刘志刚
王立有
陈隽文
韩志伟
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Southwest Jiaotong University
Original Assignee
Southwest Jiaotong University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Southwest Jiaotong University filed Critical Southwest Jiaotong University
Priority to CN201710493102.5A priority Critical patent/CN107437245B/en
Publication of CN107437245A publication Critical patent/CN107437245A/en
Application granted granted Critical
Publication of CN107437245B publication Critical patent/CN107437245B/en
Expired - Fee Related legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0004—Industrial image inspection
    • G06T7/0008—Industrial image inspection checking presence/absence
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/136—Segmentation; Edge detection involving thresholding
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/155—Segmentation; Edge detection involving morphological operators
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/70—Determining position or orientation of objects or cameras
    • G06T7/73—Determining position or orientation of objects or cameras using feature-based methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30108—Industrial image inspection
    • G06T2207/30164—Workpiece; Machine component

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Quality & Reliability (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a high-speed railway contact net fault diagnosis method based on a deep convolutional neural network, which comprises the following steps of: acquiring a two-dimensional gray image of a high-speed railway contact net supporting device; pre-training a deep convolutional neural network through a contact network training set, carrying out training in a faster RCNN, extracting an equipotential line in a two-dimensional gray image through a trained model, and segmenting to obtain an equipotential line region picture; sequentially processing the obtained equipotential line region picture, adjusting the brightness and contrast of the obtained equipotential line region picture, performing recursive Otsu pre-segmentation, performing ICM/MPM segmentation, corroding the expansion picture to obtain an equipotential line pixel point, extracting a maximum connected domain and counting the number N of independent connected domains in the equipotential line pixel point region; if N is larger than m, judging that the equipotential line parts have strand scattering faults; the method can accurately position the equipotential lines, improves the fault diagnosis accuracy, and meets the actual production requirements.

Description

High-speed railway contact net fault diagnosis method based on deep convolutional neural network
Technical Field
The invention relates to the field of high-speed railway contact network fault detection, in particular to a high-speed railway contact network fault diagnosis method based on a deep convolutional neural network.
Background
The rapid development of the high-speed railway as an important basic vehicle facility puts higher requirements on safety problems; the equipotential line is one of parts of a contact net supporting device and has the function of ensuring equipotential connection between a locator support and a locator; equipotential lines are arranged on the front side and the back side between the locator support and the locator on the high-speed railway, so that the importance of the equipotential lines can be seen; the application of the non-contact detection technology based on image processing on the railway mainly focuses on the measurement of geometrical parameters of a contact net and the detection of bad states of a pantograph-catenary, such as the detection of the inclination of a positioner, the measurement of a guide value higher than a pull-out value, the detection of wind deflection of the contact net, the detection of cracks of a pantograph slide plate and the like; aiming at the fault detection of parts of a contact net supporting device, the parts of the contact net are positioned by adopting a traditional characteristic extraction method; because the equipotential lines are non-rigid and have more shapes, the traditional template matching method cannot achieve satisfactory effects by utilizing the existing HOG characteristics or SIGT characteristics.
Disclosure of Invention
The invention provides a high-speed railway contact net equipotential line strand fault diagnosis method based on a deep convolutional neural network, which has higher detection precision.
The technical scheme adopted by the invention is as follows: the method for diagnosing the faults of the high-speed railway contact network based on the deep convolutional neural network comprises the following steps of:
acquiring a two-dimensional gray image of a high-speed railway contact net supporting device;
pre-training a deep convolutional neural network through a contact net training set, carrying the deep convolutional neural network into a target detection framework false RCNN for training, extracting and segmenting an equipotential line in a two-dimensional gray image through a trained model, and obtaining an equipotential line area picture;
sequentially processing the obtained equipotential line region picture, adjusting the brightness and contrast of the obtained equipotential line region picture, pre-dividing the obtained equipotential line region picture by an Otsu method of a recursive maximum inter-class variance method, dividing and corroding the expanded picture by using an iterative model/maximum edge posterior probability algorithm (ICM/MPM) algorithm to obtain equipotential line pixel points, extracting a maximum connected domain and counting the number N of independent connected domains in the equipotential line pixel point region;
if N is larger than m, the equipotential line parts are judged to have strand scattering faults.
Further, if N is less than or equal to m, the potential fault possibility is obtained through the standard difference value of the pixel values of the equipotential lines after the normalization formula;
the normalization method is as follows:
Figure BDA0001331952520000011
Figure BDA0001331952520000021
the standard deviation σ is calculated as follows:
Figure BDA0001331952520000022
in the formula: w is aiThe ith position pixel value of the equipotential line; w is aminIs the minimum value among the equipotential pixel values; w is amaxIs the maximum value among the equipotential pixel values;
Figure BDA0001331952520000023
the normalized equipotential line image mean value is obtained; v. ofiThe normalized pixel value of the ith position of the equipotential line is obtained; n is the number of equipotential line pixel points, and sigma is the standard deviation of the normalized equipotential pixels.
Further, the method for pre-dividing the maximum inter-class variance method Otsu method comprises the following steps:
obtaining an equipotential line area picture;
calculating a gray level histogram in the picture;
calculating the probability of occurrence of each pixel value;
traversing each pixel and calculating the inter-class variance;
and acquiring a corresponding pixel value when the inter-class variance is maximum.
Further, the ICM/MPM algorithm comprises the steps of:
according to the Bayesian formula:
Figure BDA0001331952520000024
in the formula: theta is a model parameter matrix, and y and x are sample data of the observation field and the label field respectively; p (x | y, θ) is the conditional probability of the observation field to the label field; p (y) is the prior probability of the observed field, which is a constant; converting the image segmentation problem into an optimization problem according to an MPM criterion;
Figure BDA0001331952520000025
Figure BDA0001331952520000026
Figure BDA0001331952520000027
U(x)＝∑c∈CVc(xc) (5)
Figure BDA0001331952520000028
Figure BDA0001331952520000031
in the formula:
Figure BDA0001331952520000032
for the optimized objective equation, Z is a normalized constant, C is the set of all potential masses, u (x) is the sum of potential energies of all potential masses within the set of potential masses, and T is a constant, usually taking 1, which controls the shape of p (x), the larger T, the more gradual the shape of p (x) becomes. Vc(xc) Is potential energy of potential group, S is the S-th position of picture, ysFor the pixel values of the S-position observation picture,
Figure BDA0001331952520000033
to possess a reference number xsThe average of all the pixel points of (a),
Figure BDA0001331952520000034
to possess a reference number xsThe variance of all pixel points;
the iteration is carried out by using an ICM algorithm, and the specific process is as follows:
Figure BDA0001331952520000035
Figure BDA0001331952520000036
Figure BDA0001331952520000037
in the formula: in the formula: mu.sk(p) mean of the pixels belonging to the kth class in the p-th iteration of the observation field, σk(p) variance of pixel points belonging to the kth class at the p-th iteration of the observation field, NkAnd (p) is the number of pixels belonging to the kth class in the p-th iteration of the observation field, k is the class number of image segmentation, and p is the p-th iteration.
Furthermore, the deep convolutional neural network comprises six convolutional layers, two pooling layers and two full-connection layers, wherein the back of the first two convolutional layers is connected with one down-sampling pooling layer, the back of the second pooling layer is connected with 4 convolutional layers, the four convolutional layers are connected in sequence, the back of the 6 th convolutional layer is connected with 2 full-connection layers, and the last full-connection layer outputs a 1000 x 1 vector.
The invention has the beneficial effects that:
(1) the invention adopts the deep neural network to automatically learn and extract the characteristics, so that the equipotential lines can be more accurately positioned;
(2) according to the invention, Otsu is adopted to pre-divide the picture, and then the dividing result of Otsu is optimized by adopting ICM/MPM, so that the picture dividing result is more accurate, and the accuracy of fault diagnosis is improved;
(3) the invention can give the possibility of potential failure when the failure can not be identified, and meets the requirement of actual production.
Drawings
FIG. 1 is a schematic flow chart of the present invention.
FIG. 2 is a diagram of the fast RCNN architecture of the present invention.
FIG. 3 is a diagram of CATENARNET convolutional neural network architecture in the present invention.
FIG. 4 is a flow chart of the ICM/MPM algorithm of the present invention.
FIG. 5 is a flow chart of the Otsu method of the present invention.
FIG. 6 is a diagram illustrating the effect of equipotential line splitting according to the present invention.
Detailed Description
The invention is further described with reference to the following figures and specific embodiments.
As shown in fig. 1 to 5, a method for diagnosing a fault of a high-speed railway contact network based on a deep convolutional neural network comprises the following steps:
acquiring a two-dimensional gray image of a high-speed railway contact net supporting device;
the deep convolutional neural network CATENARYNET, CATENARYNET is pre-trained by a contact network training set, is a convolutional neural network architecture designed for an equipotential line picture of a contact network, and the architecture design consists of six convolutional layers, two pooling layers and two full-connection layers, wherein the back of the first two convolutional layers is respectively connected with one downsampling pooling layer, the back of the second pooling layer is connected with 4 convolutional layers, the four convolutional layers are sequentially connected, the back of the 6 th convolutional layer is connected with 2 full-connection layers, and the last full-connection layer outputs a 1000 x 1 vector; carrying CATENARYNET into a fast RCNN for training, extracting an equipotential line in a two-dimensional gray image through a trained model, and segmenting to obtain an equipotential line region picture;
sequentially processing the obtained equipotential line region picture, adjusting the brightness and contrast of the obtained equipotential line region picture, performing pre-segmentation by a recursive Otsu method, initializing parameters, segmenting by an ICM/MPM algorithm, corroding the expanded picture to obtain equipotential line pixel points, extracting a maximum connected domain and counting the number N of independent connected domains in the equipotential line pixel point region;
if N is larger than m, the equipotential line parts are judged to have strand scattering faults.
Further, if N is less than or equal to m, the potential fault possibility is obtained through the standard difference value of the pixel values of the equipotential lines after the normalization formula; there is no definite criterion as to whether there is a stray stock; the method considers that only scatter is considered as a fault, and the rest of the conditions with the tendency of scatter are considered as the possibility of potential strand scattering faults; this probability is given by the standard deviation of the pixel values of the normalized equipotential lines; when the equipotential lines are scattered, the method judges whether the strands are scattered or not by detecting the number of independent connected domains in the area; when the strands of the equipotential lines are scattered, the place where each strand is scattered is processed to become an independent connected domain, and in practical application, only one strand of the equipotential lines is not scattered; therefore, in the method, the number m of the independent connected domains is larger than 3 and is generally used as a criterion for the divergence of the equipotential lines;
the normalization method is as follows:
Figure BDA0001331952520000041
Figure BDA0001331952520000042
the standard deviation σ is calculated as follows:
Figure BDA0001331952520000051
in the formula: w is aiThe ith position pixel value of the equipotential line; w is aminIs the minimum value among the equipotential pixel values; w is amaxIs the maximum value among the equipotential pixel values;
Figure BDA0001331952520000052
the normalized equipotential line image mean value is obtained; v. ofiThe normalized pixel value of the ith position of the equipotential line is obtained; n is an equipotential line pixelAnd the number and sigma are normalized pixel standard deviations.
Further, the Otsu method pre-segmentation method comprises the following steps:
obtaining an equipotential line area picture;
calculating a gray level histogram in the picture;
calculating the probability of occurrence of each pixel value;
traversing each pixel and calculating the inter-class variance;
and acquiring a corresponding pixel value when the inter-class variance is maximum.
Further, the ICM/MPM algorithm comprises the steps of: the ICM/MPM algorithm comprises two parts, wherein the first part converts an equipotential line segmentation problem into an optimization problem by using an MPM criterion; the second part is to solve the model to be optimized by using an ICM algorithm;
according to the Bayesian formula:
Figure BDA0001331952520000053
in the formula: theta is a model parameter matrix, and y and x are sample data of the observation field and the label field respectively; p (x | y, θ) is the conditional probability of the observation field to the label field; p (y) is the prior probability of the observed field, which is a constant; converting the image segmentation problem into an optimization problem according to an MPM (Multi-Point Measure) criterion;
Figure BDA0001331952520000054
Figure BDA0001331952520000055
Figure BDA0001331952520000056
U(x)＝∑c∈CVc(xc) (5)
Figure BDA0001331952520000057
Figure BDA0001331952520000061
in the formula:
Figure BDA0001331952520000062
for the objective equation to be optimized, Z is a normalization constant, C is a set of all potential masses, U (x) is the sum of potential energies of all potential masses in the set of potential masses, T is a constant which is usually 1, the shape of P (x) is controlled to be more gentle, and the shape of P (x) is more gentle when T is larger, and V isc(xc) Is potential energy of potential group, S is the S-th position of picture, ysFor the pixel values of the S-position observation picture,
Figure BDA0001331952520000063
to possess a reference number xsThe average of all the pixel points of (a),
Figure BDA0001331952520000064
to possess a reference number xsThe variance of all pixel points;
the iteration is carried out by using an ICM algorithm, and the specific process is as follows:
Figure BDA0001331952520000065
Figure BDA0001331952520000066
Figure BDA0001331952520000067
in the formula: mu.sk(p) mean of the pixels belonging to the kth class in the p-th iteration of the observation field, σk(p) variance of pixel points belonging to the kth class at the p-th iteration of the observation field, NkAnd (p) is the number of pixels belonging to the kth class in the p-th iteration of the observation field, k is the class number of image segmentation, and p is the p-th iteration.
Furthermore, the deep convolutional neural network comprises six convolutional layers, two pooling layers and two full-connection layers, wherein the back of the first two convolutional layers is respectively connected with a down-sampling pooling layer, the back of the second pooling layer is connected with 4 convolutional layers, the four convolutional layers are sequentially connected, the back of the 6 th convolutional layer is connected with 2 full-connection layers, and the last full-connection layer outputs a 1000 x 1 vector; each convolutional layer uses the Relu linear activation function.
When in use, the concrete working steps are as follows:
(1) continuously shooting the front side, the whole body and the local part of the contact net supporting device through a contact net suspension supporting device state detection system arranged on a patrol car, and acquiring a two-dimensional gray image of the contact net supporting device in real time; selecting pictures of the equipotential line component area of the contact net from the obtained picture frame, and manufacturing a training data set and a testing data set according to a VOC2007 standard; meanwhile, common contact net parts are cut out to be made into a contact net training set for pre-training CATENARYNET;
(2) in order to improve the positioning precision, the invention provides a new equipotential line recognition neural network structure CATENARYNET, which is more suitable for feature extraction of a catenary picture; CATENARYNET consists of six convolutional layers and two full-link layers;
(3) pre-training CATENARYNET on a contact net data set, training the contact net data set in a faster RCNN, and extracting and dividing the position of the contact net equipotential line in a two-dimensional gray picture through a trained model; adjusting brightness and contrast of the segmented equipotential line region picture, based on an Ostu algorithm segmentation result, primarily segmenting the picture by utilizing an ICM/MPM algorithm, segmenting accurate equipotential line pixel points by morphological processing of a digital image (corrosion expansion operation is used in the invention), and counting the number of independent connected domains in the equipotential line pixel point region;
the ICM/MPM algorithm is as follows:
Figure BDA0001331952520000071
in the formula: theta is a model parameter matrix, and y and x are sample data of the observation field and the label field respectively; p (x | y, θ) is the conditional probability of the observation field to the label field; p (y) is the prior probability of the observed field;
Figure BDA0001331952520000072
Figure BDA0001331952520000073
Figure BDA0001331952520000074
U(x)＝∑c∈CVc(xc) (5)
Figure BDA0001331952520000075
Figure BDA0001331952520000076
in the formula:
Figure BDA0001331952520000077
for the optimized objective equation, Z is a normalized constant, C is a set of all potential masses, U (x) is the sum of potential energies of all potential masses in the set of potential masses, T controls the shape of P (x), and the larger T, the more gradual the shape of P (x) becomes. Vc(xc) Is potential energy of potential group, S is the S-th position of picture, ysFor the pixel values of the S-position observation picture,
Figure BDA0001331952520000078
to possess a reference number xsThe average of all the pixel points of (a),
Figure BDA0001331952520000079
to possess a reference number xsThe variance of all pixel points.
The iteration is carried out by using an ICM algorithm, and the specific process is as follows:
Figure BDA00013319525200000710
Figure BDA00013319525200000711
Figure BDA0001331952520000081
in the formula: mu.sk(p) mean of the pixels belonging to the kth class in the p-th iteration of the observation field, σk(p) variance of pixel points belonging to the kth class at the p-th iteration of the observation field, NkAnd (p) is the number of pixels belonging to the kth class in the p-th iteration of the observation field, k is the class number of image segmentation, and p is the p-th iteration.
(4) And according to the statistical rule of the pixel gray values of the pixels of the equipotential lines and the number of independent connected domains in the equipotential line area, giving the potential fault possibility and the fault state of the scattered strands of the components of the equipotential lines.
Examples
The state detection monitoring device of the overhead line system suspension supporting device comprises a high-definition industrial camera, an integrated large-scale light source array, a trigger control function module, a high-performance server and the like; the high-definition industrial camera and the integrated large-scale light source array are arranged on the top of the car, when the inspection car runs on a line at a certain speed, the equipment shoots the front and back sides, the whole body and the local part of the contact net supporting device, and correspondingly stores the position information of pictures; the method mainly comprises three stages: the first stage is an equipotential line positioning stage; the second stage is an equipotential line segmentation stage; the third stage is a fault identification stage; acquiring an input picture from a state detection monitoring device of a contact net suspension supporting device; positioning an input picture by adopting a trained fast RCNN model and dividing the positioned picture of the rectangular frame; the divided picture is divided by adopting an ICM/MPM algorithm to obtain pixels of the equipotential lines and the areas of the equipotential lines in the picture; judging faults and calculating the potential fault possibility by detecting the number of independent connected domains in the region and the standard deviation of the normalized equipotential pixels; table 1 is a CAENARYNET detailed parameter configuration table:
table 1 CAENARYNET detailed parameter set-up
Figure BDA0001331952520000082
Because the size of a single channel of a picture shot by a contact net is 6600 multiplied by 4400 pixels, the parameter of the convolution layer of the first layer of the deep convolutional neural network structure is set to be 1 multiplied by 660 multiplied by 440; in order to shorten the pre-training time and not influence the classification precision, a sample with one tenth of resolution ratio is adopted for pre-training during training; still use 6600 x 4400 pixel's picture when testing, this can improve the test precision.
FIG. 2 is a family RCNN architecture diagram, during actual training, introducing the trained convolutional layer parameters of the feature extraction network in FIG. 3 into the feature extraction network in FIG. 2 for training; the fast RCNN mainly comprises two parts, wherein the first part is a risk coefficient RPN and is used as a body region candidate frame; the second part is fast RCNN; and the fast RCNN generates an ROI feature vector through the ROI pooling layer according to the suggestion of the region candidate frame and the features of the feature extraction network and respectively sends the ROI feature vector into the Softmax layer and the Bbox regression device.
The equipotential line is positioned, contrast and brightness of the cut picture are adjusted, the adjusted picture is pre-segmented by using a recursive Otsu method, then the segmentation result is corrected by adopting an ICM/MPM algorithm, then the corrected result is corroded and expanded, the maximum connected domain is extracted to serve as the final segmentation result of the equipotential line, and the core idea of the ICM/MPM algorithm is that the misclassification result of Otsu is corrected, so that the accuracy of pixel point classification is improved.
The black holes in (j) and (k) in FIG. 6 are independent connected domains; when the scattered strand fault detection is carried out, the number of the black holes needs to be detected; by adopting the fault detection method provided by the invention, the fault diagnosis is respectively carried out on the images in the figure 6 to obtain a table 2. It can be seen from table 2 that the method of the present invention can accurately determine that (a) (i) (j) (k) (l) (m) has a strand loosing fault, and (b) (c) (d) (e) (f) (g) (h) (n) (o) has no fault, and based on the potential fault probability, (g) is most likely to have a strand loosing, and (b) is least likely to have a strand loosing fault, which meets the artificial observation result of the picture.
TABLE 2 Fault diagnosis results
Figure BDA0001331952520000091
The method utilizes the deep convolution neural network to replace the traditional manual related characteristics to automatically learn and extract the characteristics, and simultaneously utilizes the neural network regression to generate the location of the regional candidate frame to replace the traditional sliding window; the equipotential line can be positioned more accurately, and the positioning time is shortened; optimizing the Otsu segmentation results using ICM/MPM after the Otsu-based segmentation results; the segmentation result is more accurate, the error rate of the segmentation result is effectively reduced, and the accuracy of fault diagnosis of the system is greatly improved; the fault is identified by utilizing the self characteristics of the equipotential line when the fault occurs, and the potential fault possibility is given when the fault cannot be identified, so that the design is more humanized, the fault can be prevented in advance, and the humanized requirement of the system in practical application is fully considered; potential fault possibility is given when the equipotential line is not in fault, and the requirement of actual production is better met; the deep convolutional neural network CATENARYNET achieves better compromise on a deep network and a shallow network, and parameter setting is more suitable for contact network pictures; the training time is shortened, the classification accuracy is improved, and the positioning accuracy is improved; the method can automatically learn the characteristics of the equipotential lines without using the characteristics of artificial design, shortens the positioning time of the equipotential lines, improves the positioning precision, improves the accuracy of fault diagnosis, can provide the possibility of potential faults, and is more suitable for practical application.

Claims (4)

1. A high-speed railway contact network fault diagnosis method based on a deep convolutional neural network is characterized by comprising the following steps:
acquiring a two-dimensional gray image of a high-speed railway contact net supporting device;
pre-training a deep convolutional neural network through a contact net training set, carrying the deep convolutional neural network into a target detection framework false RCNN for training, extracting and segmenting an equipotential line in a two-dimensional gray image through a trained model, and obtaining an equipotential line area picture;
sequentially processing the obtained equipotential line region picture, adjusting the brightness and contrast of the obtained equipotential line region picture, performing Otsu pre-segmentation by a recursive maximum inter-class variance method, performing ICM/MPM segmentation by using a conditional iteration model-maximum edge posterior probability algorithm, corroding the expanded picture to obtain an equipotential line pixel point, extracting a maximum connected domain and counting the number N of independent connected domains in the equipotential pixel point region;
if N is larger than m, judging that the equipotential line parts have strand scattering faults;
if N is less than or equal to m, the potential fault possibility is obtained through the standard difference value of the pixel values of the equipotential lines after the normalization formula;
the normalization method is as follows:
Figure FDA0002424336480000011
Figure FDA0002424336480000012
the standard deviation σ is calculated as follows:
Figure FDA0002424336480000013
in the formula: w is aiThe ith position pixel value of the equipotential line; w is aminIs the minimum value among the equipotential pixel values; w is amaxIs the maximum value among the equipotential pixel values;
Figure FDA0002424336480000014
the normalized equipotential line image mean value is obtained; v. ofiIs the pixel value of the ith position of the normalized equipotential line(ii) a N is the number of equipotential line pixel points, and sigma is the standard deviation of the normalized equipotential pixels.
2. The method for diagnosing the faults of the high-speed railway contact network based on the deep convolutional neural network as claimed in claim 1, wherein the method for pre-dividing the maximum inter-class variance method Otsu comprises the following steps:
obtaining an equipotential line area picture;
calculating a gray level histogram in the picture;
calculating the probability of occurrence of each pixel value;
traversing each pixel and calculating the inter-class variance;
and acquiring a corresponding pixel value when the inter-class variance is maximum.
3. The method for diagnosing the faults of the high-speed railway contact network based on the deep convolutional neural network as claimed in claim 1, wherein the ICM/MPM algorithm comprises the following steps:
Figure FDA0002424336480000021
in the formula: theta is a model parameter matrix, and y and x are sample data of the observation field and the label field respectively; p (x | y, θ) is the conditional probability of the observation field to the label field; p (y) is the prior probability of the observed field;
Figure FDA0002424336480000022
Figure FDA0002424336480000023
Figure FDA0002424336480000024
U(x)＝∑c∈CVc(xc) (5)
Figure FDA0002424336480000025
Figure FDA0002424336480000026
in the formula:
Figure FDA0002424336480000027
for the optimized objective equation, Z is a normalization constant, C is a set of all potential masses, U (x) is the sum of potential energies of all potential masses in the set of potential masses, T is a constant, Vc(xc) Is potential energy of potential group, S is the S-th position of picture, ysFor the pixel values of the S-position observation picture,
Figure FDA0002424336480000028
to possess a reference number xsThe average of all the pixel points of (a),
Figure FDA0002424336480000029
to possess a reference number xsThe variance of all pixel points;
the iteration is carried out by using an ICM algorithm, and the specific process is as follows:
Figure FDA00024243364800000210
Figure FDA00024243364800000211
Figure FDA00024243364800000212
in the formula: mu.sk(p) mean of the pixels belonging to the kth class in the p-th iteration of the observation field, σk(p) variance of pixel points belonging to the kth class at the p-th iteration of the observation field, Nk(p) pixel points belonging to the kth class in the p-th iteration of the observation fieldK is the number of classes of image segmentation and p is the p-th iteration.
4. The method for diagnosing the faults of the high-speed railway contact networks based on the deep convolutional neural network as claimed in claim 1, wherein the deep convolutional neural network comprises six convolutional layers, two pooling layers and two full-connection layers, the downsampling pooling layer is respectively connected behind the first two convolutional layers, the 4 convolutional layers are connected behind the second pooling layer, the four convolutional layers are sequentially connected, the 2 full-connection layers are connected behind the 6 th convolutional layer, and the final full-connection layer outputs a 1000 x 1 vector.
CN201710493102.5A 2017-06-26 2017-06-26 High-speed railway contact net fault diagnosis method based on deep convolutional neural network Expired - Fee Related CN107437245B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201710493102.5A CN107437245B (en) 2017-06-26 2017-06-26 High-speed railway contact net fault diagnosis method based on deep convolutional neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201710493102.5A CN107437245B (en) 2017-06-26 2017-06-26 High-speed railway contact net fault diagnosis method based on deep convolutional neural network

Publications (2)

Publication Number Publication Date
CN107437245A CN107437245A (en) 2017-12-05
CN107437245B true CN107437245B (en) 2020-10-20

Family

ID=60458622

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201710493102.5A Expired - Fee Related CN107437245B (en) 2017-06-26 2017-06-26 High-speed railway contact net fault diagnosis method based on deep convolutional neural network

Country Status (1)

Country Link
CN (1) CN107437245B (en)

Families Citing this family (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108387814B (en) * 2018-01-12 2021-04-13 广州市扬新技术研究有限责任公司 Urban rail transit contact network fault positioning device and method with power supply switched on two sides
CN108573277A (en) * 2018-03-12 2018-09-25 北京交通大学 A kind of pantograph carbon slide surface disease automatic recognition system and method
CN108647045B (en) * 2018-03-20 2021-10-01 科大讯飞股份有限公司 Method and device for realizing activation function, storage medium and electronic equipment
CN108664997A (en) * 2018-04-20 2018-10-16 西南交通大学 High iron catenary equipotential line defective mode detection method based on cascade Faster R-CNN
CN108828405B (en) * 2018-06-06 2020-05-12 西南交通大学 Power transmission line fault detection method based on random matrix
CN110008971B (en) 2018-08-23 2022-08-09 腾讯科技（深圳）有限公司 Image processing method, image processing device, computer-readable storage medium and computer equipment
CN109241982B (en) * 2018-09-06 2021-01-29 广西师范大学 Target detection method based on deep and shallow layer convolutional neural network
CN109360188A (en) * 2018-09-13 2019-02-19 西南交通大学 A kind of detection method of high iron catenary equipotential line defective mode
CN109636789B (en) * 2018-12-13 2021-07-30 银河水滴科技（北京）有限公司 Image segmentation-based part missing detection method and device for track fastener
CN109583573A (en) * 2018-12-13 2019-04-05 银河水滴科技（北京）有限公司 A kind of part missing detection method and device of rail clip
CN109753962B (en) * 2019-01-13 2023-06-09 南京邮电大学盐城大数据研究院有限公司 Method for processing text region in natural scene image based on hybrid network
CN110135273B (en) * 2019-04-19 2020-05-22 中铁第一勘察设计院集团有限公司 Contact network video image cloud intelligent monitoring and fault identification method
CN110084852A (en) * 2019-04-22 2019-08-02 西南交通大学 A kind of accurate positioning method of high iron catenary support device sleeve puller bolt
CN110210427B (en) * 2019-06-06 2021-04-23 中国民航科学技术研究院 Corridor bridge working state detection system and method based on image processing technology
CN110428400B (en) * 2019-07-09 2021-08-27 北京百分点科技集团股份有限公司 Method, device, equipment and storage medium for positioning equipment fault
CN110503048B (en) * 2019-08-26 2020-07-17 中铁电气化局集团有限公司 Identification system and method for suspension device of rigid contact net
CN111311591A (en) * 2020-03-07 2020-06-19 西南交通大学 Method for detecting lifting amount of high-speed railway contact net
CN111784656A (en) * 2020-06-28 2020-10-16 京东数字科技控股有限公司 Railway contact network fault detection method and device, electronic equipment and storage medium
CN112508911A (en) * 2020-12-03 2021-03-16 合肥科大智能机器人技术有限公司 Rail joint touch net suspension support component crack detection system based on inspection robot and detection method thereof
CN113658112B (en) * 2021-07-27 2024-03-01 北京交通大学 Bow net anomaly detection method based on template matching and neural network algorithm

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101247030A (en) * 2007-08-01 2008-08-20 北京深浪电子技术有限公司 Overhead network obstacle detouring inspection robot and its obstacle detouring control method
CN101577003A (en) * 2009-06-05 2009-11-11 北京航空航天大学 Image segmenting method based on improvement of intersecting visual cortical model
CN101699511A (en) * 2009-10-30 2010-04-28 深圳创维数字技术股份有限公司 Color image segmentation method and system
CN101847259A (en) * 2010-01-21 2010-09-29 西北工业大学 Infrared object segmentation method based on weighted information entropy and markov random field
CN102968637A (en) * 2012-12-20 2013-03-13 山东科技大学 Complicated background image and character division method
CN104036491A (en) * 2014-05-14 2014-09-10 西安电子科技大学 SAR image segmentation method based on area division and self-adaptive polynomial implicit model
CN105741291A (en) * 2016-01-30 2016-07-06 西南交通大学 Method for detecting faults of equipotential lines of high-speed railway overhead line system suspension devices
CN105957073A (en) * 2015-04-29 2016-09-21 国网河南省电力公司电力科学研究院 Fault detection method for scattered strands in power transmission line
CN106683099A (en) * 2016-11-17 2017-05-17 南京邮电大学 Product surface defect detection method

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101247030A (en) * 2007-08-01 2008-08-20 北京深浪电子技术有限公司 Overhead network obstacle detouring inspection robot and its obstacle detouring control method
CN101577003A (en) * 2009-06-05 2009-11-11 北京航空航天大学 Image segmenting method based on improvement of intersecting visual cortical model
CN101699511A (en) * 2009-10-30 2010-04-28 深圳创维数字技术股份有限公司 Color image segmentation method and system
CN101847259A (en) * 2010-01-21 2010-09-29 西北工业大学 Infrared object segmentation method based on weighted information entropy and markov random field
CN102968637A (en) * 2012-12-20 2013-03-13 山东科技大学 Complicated background image and character division method
CN104036491A (en) * 2014-05-14 2014-09-10 西安电子科技大学 SAR image segmentation method based on area division and self-adaptive polynomial implicit model
CN105957073A (en) * 2015-04-29 2016-09-21 国网河南省电力公司电力科学研究院 Fault detection method for scattered strands in power transmission line
CN105741291A (en) * 2016-01-30 2016-07-06 西南交通大学 Method for detecting faults of equipotential lines of high-speed railway overhead line system suspension devices
CN106683099A (en) * 2016-11-17 2017-05-17 南京邮电大学 Product surface defect detection method

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
基于Markov随机场和K均值聚类的小麦叶部病害图像分割;黄帅;《万方数据》;20170212;全文 *
基于深度学习的高铁接触网定位器检测与识别;陈东杰等;《中国科学技术大学学报》;20170430;第322-325页、图2 *
基于计算机视觉的刀具后刀面磨损检测技术;刘荣涛;《中国优秀硕士学位论文全文数据库》;20090115;第26-28、35-41、50-53页 *
输电线路远程智能巡线系统的设计与实现;王忠强;《万方数据》;20150701;第38-40页 *

Also Published As

Publication number Publication date
CN107437245A (en) 2017-12-05

Similar Documents

Publication Publication Date Title
CN107437245B (en) High-speed railway contact net fault diagnosis method based on deep convolutional neural network
CN115082683B (en) Injection molding defect detection method based on image processing
CN106875373B (en) Mobile phone screen MURA defect detection method based on convolutional neural network pruning algorithm
CN110210463B (en) Precise ROI-fast R-CNN-based radar target image detection method
CN109977808B (en) Wafer surface defect mode detection and analysis method
CN111179251B (en) Defect detection system and method based on twin neural network and by utilizing template comparison
CN111462076B (en) Full-slice digital pathological image fuzzy region detection method and system
CN105044122A (en) Copper part surface defect visual inspection system and inspection method based on semi-supervised learning model
CN108537751B (en) Thyroid ultrasound image automatic segmentation method based on radial basis function neural network
CN111611874B (en) Face mask wearing detection method based on ResNet and Canny
CN110648310B (en) Weak supervision casting defect identification method based on attention mechanism
CN113724231B (en) Industrial defect detection method based on semantic segmentation and target detection fusion model
CN112102229A (en) Intelligent industrial CT detection defect identification method based on deep learning
CN114372955A (en) Casting defect X-ray diagram automatic identification method based on improved neural network
CN109242032B (en) Target detection method based on deep learning
CN112365497A (en) High-speed target detection method and system based on Trident Net and Cascade-RCNN structures
CN112381175A (en) Circuit board identification and analysis method based on image processing
CN114926407A (en) Steel surface defect detection system based on deep learning
CN111860106A (en) Unsupervised bridge crack identification method
KR20220014805A (en) Generating training data usable for examination of a semiconductor specimen
CN114463843A (en) Multi-feature fusion fish abnormal behavior detection method based on deep learning
CN116012291A (en) Industrial part image defect detection method and system, electronic equipment and storage medium
Zhao et al. Research on detection method for the leakage of underwater pipeline by YOLOv3
CN114782410A (en) Insulator defect detection method and system based on lightweight model
CN111160100A (en) Lightweight depth model aerial photography vehicle detection method based on sample generation

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant
CF01 Termination of patent right due to non-payment of annual fee
CF01 Termination of patent right due to non-payment of annual fee

Granted publication date: 20201020

Termination date: 20210626